Does the AI model really lead the domestic "capital feast"? yes and no

Source: Internet Phantom Thieves

Image source: Generated by Unbounded AI

The AI model is the hottest term in the capital market so far in 2023, both in the United States and in China. On various social media platforms, consulting companies and self-media can be seen every day announcing that "AI large-scale models have set off a wave of investment in China", and even proclaimed convincingly that "the large-scale model track has attracted tens of billions/ Hundreds of billions of investment". The AI expos or seminars held in Shanghai and Beijing were full, and it seemed that all the analysts and fund managers were present, and they took pictures on the spot and sent them to Moments. In the eyes of many people, large models (and AIGC applications) have become a super investment track comparable to new energy and chips back then. "Capital Feast".

However, the actual situation I have seen does not support the above point of view. After in-depth discussions with familiar investors and entrepreneurs in the primary market, I found that the actual popularity of the AI big model in the so-called "capital circle" (especially the venture capital circle) is far less than its popularity on social media; investors Although they are concerned about the AI track, they largely stay in the state of "lips and bolts". As for the so-called "large-scale model track has attracted tens of billions/hundreds of billions of investment", it is even more like a dream, which is completely contrary to the facts.

In summary, what I see is:

• The hype on the "AI big model" is limited to the secondary market, to be precise, the A-share secondary market.
• In the primary market, investors are willing to learn and research AI projects, but they are very cautious in investing.
• It is extremely difficult for startups on the AI track to obtain investment, and the number of "unicorns" born is very small.
• Major Internet companies have invested a lot of internal resources in the AI model, but it is difficult to call it "investment".

Undoubtedly, in the A-share market in the first half of 2023, any listed company with big models and AIGC will be speculated. The problem is that the TMT industry in the A-share market has always been out of touch with the industry, and the transmission mechanism to the primary market is very weak, and this time is no exception. We all know that no matter how many billions of market capitalization these companies, such as 360, Kunlun Wanwei, Tom Cat, and Chinese Online, are hyped up in the A-share market, not much money will actually flow to AI large-scale models and application development itself; It can increase the divorce rate of executives of "big model concept companies".

Venture capitalists in the primary market are of course happy to learn large models, participate in AIGC exhibitions, and research AI basic R&D and application projects, but they will be very cautious when investing. Therefore, it is very difficult for start-up companies in the domestic AI track to obtain financing, whether they are in the field of large-scale model development or application. So far, only a few "unicorns" such as MiniMax and Light Year Beyond have been born on the entire track, among which Light Year Beyond also suffered a half-way failure due to Wang Huiwen's depression. In addition, even Wang Xiaochuan's Baichuan Intelligent has received quite limited investment, let alone other start-up companies without background.

Of course, the entire Internet industry has invested heavily in AI large-scale models. Many major Internet companies have listed it as a "No. 1 project", and even suspended many basic research and development projects in order to concentrate resources. The problem is that this kind of internal investment cannot be called "investment" - a large factory has allocated 10 billion yuan of resources for the recruitment of large-scale models, the purchase of hardware equipment, or the rental of cloud services. Model track investment of 10 billion", because the so-called "investment" in the capital market is an external behavior.

Therefore, the actual popularity of AI large models in the "capital circle" has been seriously overestimated. Some people are optimistic that in the next few months, the actual popularity will increase significantly, because regulatory measures for generative AI have been implemented, more talents and resources have poured in, and the development of the entire industry has become more formal. I do not think so. In fact, the domestic AI entrepreneurship track is facing two fundamental problems: the first is the problem of the industry itself, and the second is the supply problem of the domestic venture capital community. Neither problem is likely to be resolved in the short term.

Let me talk about the first one first. The basic research and development of AI large models is a highly specialized and narrow subdivision track; it is also a business that burns money and requires extremely high capital expenditures. The combination of the two has resulted in very few venture capital funds that can actually invest in large models. An important reason why MiniMax can get a higher valuation is that it was established early and has accumulated more in its history of more than two years; most of the new companies established since 2023 do not reassure investors ( Light-years away could have been the exception, but it's a pity).

As for AIGC's application of this track, its theoretical upper limit can be much higher than the development of the large model itself, but it is still largely on paper. The era when the domestic primary market spent a lot of money casually for a "consumer application" of cutting-edge technology has passed. Everyone knows that after the Internet traffic dividend is exhausted, the application of any new technology is difficult to be like 5-10 years ago. Grow up that fast. What's more, in the foreseeable future, it will be difficult for domestic consumer-grade AIGC applications to use large overseas models such as GPT, and can only be based on domestic models with lower levels, which further limits their development prospects. At least as far as the primary market is concerned, at this moment, "AIGC application" is still a track with little attention and it is difficult to get real money; I estimate that this phenomenon will continue for more than a year.

Let me talk about the second article. Since 2021, the supply-side trend of the entire domestic venture capital (VC) field has been: US dollar funds are going from bad to worse, and the source of funds for RMB funds has been shifting from "marketization" to "state-owned assets." Needless to say, the ebb tide of US dollar funds has been actively or passively shrinking the front line when the US stock market is semi-closed to domestic companies. For RMB funds, regardless of the nature of their own managers (GP), their sources of funds are increasingly dependent on local state-owned platforms, because that is currently the only relatively smooth financing channel.

Now that it has taken money from local state-owned assets, the RMB Fund must give an explanation to the local government, and the local government is most concerned about local employment and GDP. New energy, semiconductors, and biomedicine all belong to the modern manufacturing industry with a long industrial chain, which is conducive to GDP, and there must be a certain link in it that is conducive to employment; the AI model is not like this. Even if a large-scale start-up company is willing to relocate the entire R&D team to a certain province (that is impossible), the employment it creates for the local area is still negligible. As for AIGC application services, if it develops in the future, the GDP created will only belong to a few first-tier cities such as Beijing, Shanghai, Guangzhou and Shenzhen (plus Hangzhou at most), and most places have no interest in it.

If it was 5-10 years ago, the RMB fund could still be "secretly hidden" and take the money that local state-owned assets used to develop "hard technology industries such as chips" and invest it in the large-scale model track. But now, local state-owned assets are already very shrewd, and they often impose strict restrictions on investment, thus plugging all loopholes. For example, in the field of pharmaceutical investment, state-owned LPs will even require GPs to “only invest in a few very narrow categories of medical devices, and are not allowed to invest in other devices, let alone invest in tracks other than medical devices.” This is especially true for the TMT field. It is no exaggeration to say that without the approval of state-owned LPs, most RMB funds would never want to invest a penny in the AI track.

In addition to US dollar funds and RMB funds, there is another important force in the domestic venture capital industry, that is, corporate venture capital (CVC) with major Internet companies as its core. Regrettably, starting from 2021, the investment business of major Internet companies has continued to shrink; as for companies in other industries, they lack sufficient financial resources. For major Internet companies, the focus of developing generative AI is still internal, and external investment is more to play the role of "looking for spare tires". As we all know, spare tires do not need to have many, nor do they need to be given a high valuation.

Yes, the competent authorities have issued some documents to support the generative AI industry; yes, the official media is applauding large models and related applications-but these supports are still far from being able to compete with those in the fields of new energy, semiconductors, and biomedicine. Dao received support on a par. Moreover, the effect of these supports has largely fallen on the Internet giants rather than start-up companies. The benefits obtained by AI track startups may not be as good as the game industry before the second version number crisis.

The above facts should not affect the A-share market’s speculation on large models, because as we all know, the A-share market is determined by its own logic (deviating from fundamentals). Whether the AI model leads the domestic "capital market feast" depends purely on your perspective: the feast only exists in the A-share secondary market. If you exclude this market, there will be no feast at all.